[ahrazzle]

Change and scope

New package: research/pr5270-reference-repair-20261023/. It is the successor step to the audit merged as #5349, whose README named the first open step as correcting the bridge and idle transcription in a new package and comparing the corrected path against the formal controller.

The repair is a transformation of the emitted gate list, and nothing else:

1. The first 2(p−1) odd-role BRIDGE gates (the forward loop) have their endpoints swapped. Later odd bridges belong to the final reverse sweep. Their direction is already correct, so they stay as written.
2. IDLE gates tagged middle become identity, because thinning removes their even stages.
3. IDLE gates tagged initial or terminal have their power halved, because only the thick stages of each interval survive thinning.

The repaired list has the same gate count as the original. At p19 the repair touches 72 of 259 gates: 36 bridges swapped, 18 middle idles zeroed, 18 outer idles halved. In his review of the audit, @jwr404 noted that the initial idles share the terminal idle defect at p19, with nine retained. The outer correction covers both, and the p19 equality below includes all nine.

Validation

CPython 3.12, Linux, standard library only:

python3 -B research/pr5270-reference-repair-20261023/verify.py
python3 -I -O -B research/pr5270-reference-repair-20261023/verify.py

Both modes pass. For every p from 2 to 19, at the audit specialization (q=1, odd stage letters i, even stage letters 1):

• The original gate list, read by a plain interpreter with no correction flags, differs from the formal thinned controller. The mismatch count at p19 is 351, the audit's number.
• The repaired gate list, read by the same interpreter, equals the formal controller exactly. Mismatches after repair are 0 at every p.
• Leaving out any one of the three corrections breaks the equality at p=2, p=3 and p=19. All nine negative controls are rejected.

The vendor files are byte-identical to the pins in the audit package. RESULT.json records the full per-p table.

Scope limits

This is the transcription step only. Binding one actual rank-eight native reverse invocation with its independent blocks, then recalculating the complete invoice, remain open in the audit's stated order. No new κ is claimed.

Credit

The audit package and its retained lineage are credited in the package README: DaysSky, Cayden Burkart / SethBurkart123, Dillon, Dugongue / mdumitrean, Cinna Davis, Cao Yifan, zxv and Jacob Alan Sussman.

Automated posting by agentic team with human oversight.

[jwr404]

Independent review (jwr404, reviewer of #5270/#5349). Base: upstream/main 1ec1119, Python 3.15.1 locally, standard library only.

Reproduced as described.

• python3 -B verify.py passes, and so does python3 -I -O -B verify.py (0.9 s).
• The output equals RESULT.json after key-sorted JSON comparison.
• The three vendor/ files are byte-identical (sha256) to the copies in the merged research/pr5270-reference-audit package.
• The repair counts per p are 2(p−1) bridges swapped, (p−1) middle idles zeroed and (p−1) outer idles halved. At p19 that is 36+18+18 = 72 touched gates, as stated. The nine single-class negative controls fail as they should.

Stronger check: the repair is not an artefact of evaluating at i. The package compares at a single point: every kept stage letter equals i. At that point i⁴ = 1, so a wrong power could match by coincidence (e.g. "halve the outer idle power" against "keep a quarter"). I re-ran the comparison with a generic letter value z. I used the same retained controller/keep_of/thin and the same listing()/repair(). My own exact interpreter (Fractions) scales LETTER by z, scales IDLE by z^power, and evaluates the formal terms at z^(Σ exponents). The repaired list equals the formal controller entrywise for every p = 2..19 at each of these:

z ∈ { i, −1, 2, 2+i, 3−2i }

So the three corrections agree with the controller as an identity in a single shared letter variable, not just at its 4th-root-of-unity specialization. That is a meaningfully stronger statement, and the README may want to say it (script is ~50 lines; happy to share).

Remaining scope limits (correctly stated, plus two worth adding):

1. The q exponent (term[0] ∈ {±1}) is still set to 1. The gate listing has no q, so the comparison cannot test it.
2. All kept stages share one letter. A per-stage-letter comparison would need LETTER gates to carry their stage index.

Neither affects the stated claim, which is "reference transcription only".

Packaging. There is no VERIFY.json or TASK.json, which is why replay shows as skipped. scripts/check_research_commit.py --package research/pr5270-reference-repair-20261023 stops at missing or external file: …/VERIFY.json. The package is tiny and dependency-free, so registering it (python3 -B verify.py, Python 3.12, no requirements) would put it under hosted replay. Note #5386: registering also needs the full TASK.json schema (schema_version, scope_status, outcome ∈ {partial, completed, …}, open_questions, …).

Verdict: the claim is correct and stronger than stated. Register it for CI if possible. No κ is claimed or affected.